Reproduce “V-OCBF: Learning Safety Filters from Offline Data via Value-Guided Offline Control Barrier Functions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Mumuksh Tayal, Manan Tayal, Aditya Singh, Shishir Kolathaya, Ravi Prakash

V-OCBF presents a framework for learning neural Control Barrier Functions (CBFs) from offline datasets, providing strict, state-wise safety guarantees for autonomous systems. It leverages a model-free, finite-difference barrier recursion and an expectile-based objective to handle out-of-distribution actions. The system achieves 98.28% safe episodes in 3D Dubins' car collision avoidance and near-zero violations in high-dimensional MuJoCo environments, effectively synthesizing safe policies while maintaining task performance.

Key insights

  • A model-free, finite-difference barrier recursion can be derived from the continuous-time HJB-VI, enabling the learning of forward-invariant safety certificates directly from state transitions without explicit dynamics knowledge.
  • Using an expectile-based objective in offline barrier learning effectively approximates a data-supported constrained maximization, preventing extrapolation to unobserved actions and promoting stable, robust learning.
  • Decoupling the learned dynamics model—using it only for Lie derivative computation during inference and not during barrier training—is critical for robust performance, avoiding suboptimal barrier targets from potentially inaccurate dynamics predictions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.